标签: machines

A Washington Post article by the usually-interesting Joel Achenbach tackles fears many thinkers have about the future of AI. Will intelligent machines rule the world? Kill off humanity?   I remember visiting Marvin Minsky’s AI lab at MIT around 1970 while scouting colleges. They had a PDP-12 computer, by today’s standards a joke of a machine. Yet it was big and expensive. Today, a $500 iPad has more computing power than the fastest computer in the world in 1985, which cost $35 million ($77m in today’s dollars). The people interviewed in the article indirectly address the exponential growth of compute power, wondering what it bodes for the future.   Everyone talks about Moore’s Law, but few outside the industry understand it. Tom Freeman of the New York Times is completely off-base in his interpretations of it; he equates the “law” with pretty much all progress. Moore’s Law is not a law, but an observation. There is no reason to think it will continue. It has continued because it’s also an aspiration – the semiconductor people work hard to try to double transistor density every two years. That has slowed recently, and the future looks like more scaling but at increasingly unattractive prices. There’s another “law” called Dennard Scaling which says with each doubling we get all sorts of other benefits like lower power, faster clocks, etc. Few outsiders understand how much this scaling contributed to faster machines. Alas, Dennard Scaling completely fell apart at the 90 nm node. And there are other roadblocks that may slow computation progress, like the mismatch between memory and CPU speeds. Time will tell.   The article talks about increasingly-ubiquitous robots. It claims that people are better than robots at fine, agile motions, but that isn’t true. Today’s electronics cannot be assembled by humans – humans can’t even place the components on a circuit board due to the level of precision needed. A pick and place machine is needed for that. To make an IC a stepper has to position masks to a superhuman precision. So robots can be quite agile.   But today’s robots are amazingly dumb. They can only do simple, repetitive actions. There’s a lot of work going on to make them smarter, and I have no doubt that we will see robots that can put away the dishes, do the laundry, etc., probably in my lifetime. Certainly in my kids’ lifetimes.   We will see driverless cars quite soon and they will probably be better drivers than we are. But these technologies bring up the central idea of the article (which is only barely alluded to): that of machines making ethical choices. Today there’s a lot of thought going into problems that go beyond the technical. For instance, should a car break the law to increase safety? It might be impossible to merge onto a high speed road with human drivers going bumper to bumper at 85 MPH. A robot car that stuck to the speed limit will never make the merge.   Or suppose there’s a situation where the only choices are to run though a crowd of people or plow over a baby carriage. In the real world this choice never materializes as a human driver panics and takes some random action. But the people designing a driverless car’s software have to program these decisions into the code. These are rather philosophical and legal issues. Philosophers and legislators, though, are abdicating the thinking to the techies. I’m not sure that is healthy.   Ray Kurzweil is mentioned in the article. He has done some interesting work and predicts that computers will have the power of the human brain by about 2030-2040. He goes on to make some interesting, but I think intellectually-suspect, arguments about the implications. It’s not at all clear to me that lots of compute logic is the same as thinking. It might be, but no one really knows.   While the article goes on to blue-sky about super-intelligent computers wreaking havoc, I think a more likely threat, and much more imminent, is the destruction of work by smart machines. Driverless cars will mean the end of cabbies, truck drivers, garbage collectors, UPS delivery people, and far more. Automation is changing manufacturing. Even Foxconn in China is installing 1 million robots to replace workers there. When Chinese workers are too expensive, what happens to those trying to compete in the west? Even legal research is in some cases done by computers today. Maybe receptionists will be replaced , too. Grocery stores now have automated checkouts. It’s hard to imagine many jobs, other than those in science, engineering, etc., that can’t be automated.   So what happens when the machines replace, say, 20% of the workforce? I expect to see that in a decade or two. 40%? 60%?   When robots can make everything, including robots, and mine raw materials, labor will have little value. Will everything then be free?   Or will we see a Marxian situation: the mass of workers, no longer able to make a living, and the owners of the mines and robots somehow exploiting the former worker class? There’s a lot of vapid thinking today about the “1% vs the 99%”, but more than a little of income inequality is a symptom of automation.   How about a Luddite-like revolution, but one of national or even global scope? Yet one can’t really stop the march of technology. In “A Canticle for Leibowitz” nuclear war destroys civilization. The people hold techies accountable and kill all except Leibowitz, an engineer, who founds a monastery to preserve scientific knowledge. A thousand years later the tech is back, but people haven’t changed so the same problem repeats itself.   Perhaps there will be a Star Trek-like economy where everything is free, there are no wages, so people pursue activities to better themselves. I don’t have enough faith in human nature to think that’s likely. Seems to me that people will still be people, and idle hands are not a good situation.   I worry about the destruction of work and the effect it will have on society. Ironically, it’s we, the readers, who build the machines that replace labor. I feel it’s foolish to try and halt technological progress, as that also holds the only hopes we have to deal with other huge and looming problems. What’s the answer? Do you have any thoughts?

　　据WSJ报道，一家神秘的芯片初创企业Soft Machines刚刚揭开了其神秘的面纱，它的目标很有野心：要实现内核的虚拟化。   　　这家初创企业的名字叫做Soft Machines，总部位于硅谷，由英特尔前雇员Lingareddy和Mohammad Abdallah联合创立于2007年。目前Soft Machines共有250名员工，在印度和俄罗斯设有分支机构。此前这家芯片公司一直处于隐身模式，本周四，这家公司首次现身，在芯片业研究机构Linley Group举办的活动上公布了自己的计划。   　　我们知道，芯片的工作频率(时钟频率)1990年代及2000年代早期一直在稳步提升，但是主频太快会导致芯片出现功耗过大和过热的问题，因此英特尔等芯片制造商开始走多核化的路线，即限制单个微处理器的主频，通过集成多个处理器内核来提高处理性能。 这属于一种分布式分而治之并发处理的思路 ，云计算、云存储、分布式网络等等都是用这种思路来解决规模问题。   　　问题是在应用端，能充分利用多核处理优势的寥寥，所以给用户带来的速度提升感知越来越不明显。Soft Machines决心要改变这种状况。 其基本思路也是一样—分而治之，把计算任务拆分为可并发运行的更小部分。但是Soft Machines的做法有所不同。   　　以往，芯片要程序员设计产品来发送独立的指令流(即所谓的线程)，然后由处理器芯片内的各个内核进行处理，也就是说，任务的分解需要应用开发者来设计实现。这无疑提高了充分发挥多核处理器性能的门槛。   　　而现在，Soft Machines开发了一种特殊的电路模块，这种模块可以自动将线程分解，然后传递给所谓的虚拟内核，再由这些处理引擎对任务进行分配(虚拟硬件线程)。   　　Soft Machines把这种新型的CPU架构称为是VISC，以区别于以往的CISC和RISC架构。VISC可以基于不同的应用需求动态分配资源，对单/多线程的应用在性能与功耗之间做出平衡。这种方式比传统的内核调度更加灵活，效率更高，而且省却了开发者的干预。根据Soft Machines对其芯片工作样本的测试，其计算性能是普通多核处理器的2到4倍。Soft Machines据称拥有微芯片方面的30项专利。   　　这意味着对芯片的设计可以有多种选择：即可让它保持正常时钟频率下获得显著的性能提高，也可以让芯片维持在较早前的性能水平，但是却因此大幅降低功耗，从而提高了电池的续航时间。   　　 不过，业界对此可能会提出质疑。因为早在1990年代时即有人进行过芯片虚拟化的努力。 当时一家名为Transmeta的初创企业也想把软件从芯片硬件中抽象出来，但在秘密攻关数年后仍宣告失败。   　　 但是Soft Machines除了技术上不一样以外，它的商业策略也有所不同。 它不打算自己生产芯片，而是要把自己的发明卖给芯片公司，这样一来可以显著降低变现成本。Soft Machines把初始客户定位在Android芯片生产商上，尽管这个领域ARM占据了绝对的主导地位，但是Soft Machines在这个蓬勃发展的生态体系中应该仍能觅得不少商机。此外，Soft Machines称自己的技术也可以运行为英特尔、IBM等其他公司的芯片编写的软件。这也许就是Soft Machines之所以得名的原因，而这种开放性也可以给它带来更加广阔的市场空间。   　　迄今为止Soft Machines似乎赢得了不少人的青睐。融资额说明了一切。上个月底，CBinsights曾做出了一个隐身模式初创企业融资排行榜，Soft Machines以9600万美元高居榜首。而根据演示的片子，现在其总融资额已达1.25亿美元，由此可推断最近一个月它又获得了将近3000万美元的融资。投资者当中包括了Albert Yu(虞有澄)和Richard Wirt这两位英特尔的前高管，Gordon Campbell，以及三星的风投机构、AMD还有阿联酋的投资机构Mubadala。 （ WSJ）     关注集芯城微信号：icjxc520，最新IC采销行情、新品发布、电子干货、商情资讯应有尽有！